What the documents look like
Channel diversion work concerns regional and price order. Anti-counterfeit code records trace where goods have moved, e-commerce platform prices support comparison against low-priced storefronts, distributor shipment records help verify the origin of goods, and market inspections with investigation photos add on-site detail. In store service and internal training, these materials enter the knowledge base as diversion complaint forms, penalty notices, anti-counterfeit code records and investigation photos, and are retrieved as cited content. The brief states a weekly and irregular update frequency. Retrieval operates on what has actually been ingested; the update rhythm does not by itself widen or narrow the recall scope. The subject, region, time and status fields in these materials are what later determinations about whether a diversion fact holds depend on.
Constraints at the retrieval step
The knowledge base search node can associate one or more knowledge bases that use the same vector model. Bases built on different vector models cannot be associated together; the vector model identifier should be checked first and non-matching bases excluded, without running cross-model retrieval to compare missed recall. Search runs at the level of the whole knowledge base, and a collection is only a classification layer for data, unrelated to search results. Semantic retrieval relies on vector distance, full-text retrieval relies on keyword hits, and hybrid retrieval merges the two paths through RRF. The minimum relevance value takes effect only under semantic retrieval or when result reranking is used; it is a filter threshold and does not perform permission or entity isolation. The citation limit is measured in tokens and controls the length of returned content. Question optimization performs coreference resolution and question expansion and adds an extra model call. Image retrieval requires image input and the corresponding model capability, and reranking requires a rerank model enabled by configuration. The output citation content is returned as an array, and its length can be 0.
Configuration and implementation checks
Hypothetical drill sample: a brand channel control department receives a diversion complaint form whose text reads "Complaint subject: East China distributor A; goods involved: batch B2024; complaint reason: a sales link below the suggested retail price was found on an e-commerce platform; complaint date: 3 June 2024; attachments: anti-counterfeit code records and investigation photos." A draft sentence is also under review: "Distributor A has completed enforcement of the diversion penalty." The sample does not provide a penalty notice, and the enforcement status is not stated.
| Configuration item | Hypothetical trial starting point for this page | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate one diversion review base that uses the same vector model | Check each base's vector model identifier first and exclude inconsistent ones |
| Search mode | Test hybrid retrieval first | The complaint form contains a batch number and natural-language description; compare the semantic and full-text recall paths |
| Citation limit | Set 2000 tokens as a trial starting point | Test content completeness with complaint forms and penalty notices of different lengths, and compare against a shorter limit |
| Minimum relevance | Set 0.5 as a trial starting point | Label relevant and irrelevant samples first, then observe how filtered citations change |
| Question optimization | Off at first, enabled only for multi-turn follow-ups | Compare the original question with the optimized question and check whether the subject was altered |
| Result reranking | Enabled when a rerank model is available | Compare candidate ordering before and after enabling, and observe the score-based filtering |
Three common errors
- Treating the minimum relevance value as a permission or entity isolation control. That value only filters low-relevance results; it does not exclude unrelated businesses or unauthorized records. Retrieving a piece of content does not mean that content may be used to enforce a penalty.
- Treating retrieval results as business authorization. Knowledge base search returns cited content and does not constitute approval of a distributor penalty, a refund or a contract termination. Penalty enforcement requires a penalty notice and the corresponding authorization record.
- Treating a pending-review status as completed. The draft sentence in the sample states that diversion penalty enforcement has been completed, but the penalty notice is not provided and the enforcement status is not stated, so the sentence is unsupported and the evidence gap should be recorded.
How to verify the setup
- Input the sample text and the draft sentence, then compare the draft sentence's claim of completed penalty enforcement against the "complaint reason" and "complaint date" fields in the sample. The sample does not state a penalty notice or an enforcement status, so the draft sentence is unsupported and should be recorded as pending verification. The condition is met only if retrieval returns the original penalty notice and that notice states enforcement is complete.
- Input an explicit paper empty result [] and check whether the downstream design handles it as empty citation content. The retrieval output length can be 0; this does not claim that a fabricated rule makes retrieval always return [].
- Check the vector model identifier of each associated knowledge base and compare whether the identifiers match. After excluding non-matching bases, record the actual observation, and do not run cross-model retrieval to compare missed recall.